0001 =====================
0002 Overcommit Accounting
0003 =====================
0004
0005 The Linux kernel supports the following overcommit handling modes
0006
0007 0
0008 Heuristic overcommit handling. Obvious overcommits of address
0009 space are refused. Used for a typical system. It ensures a
0010 seriously wild allocation fails while allowing overcommit to
0011 reduce swap usage. root is allowed to allocate slightly more
0012 memory in this mode. This is the default.
0013
0014 1
0015 Always overcommit. Appropriate for some scientific
0016 applications. Classic example is code using sparse arrays and
0017 just relying on the virtual memory consisting almost entirely
0018 of zero pages.
0019
0020 2
0021 Don't overcommit. The total address space commit for the
0022 system is not permitted to exceed swap + a configurable amount
0023 (default is 50%) of physical RAM. Depending on the amount you
0024 use, in most situations this means a process will not be
0025 killed while accessing pages but will receive errors on memory
0026 allocation as appropriate.
0027
0028 Useful for applications that want to guarantee their memory
0029 allocations will be available in the future without having to
0030 initialize every page.
0031
0032 The overcommit policy is set via the sysctl ``vm.overcommit_memory``.
0033
0034 The overcommit amount can be set via ``vm.overcommit_ratio`` (percentage)
0035 or ``vm.overcommit_kbytes`` (absolute value). These only have an effect
0036 when ``vm.overcommit_memory`` is set to 2.
0037
0038 The current overcommit limit and amount committed are viewable in
0039 ``/proc/meminfo`` as CommitLimit and Committed_AS respectively.
0040
0041 Gotchas
0042 =======
0043
0044 The C language stack growth does an implicit mremap. If you want absolute
0045 guarantees and run close to the edge you MUST mmap your stack for the
0046 largest size you think you will need. For typical stack usage this does
0047 not matter much but it's a corner case if you really really care
0048
0049 In mode 2 the MAP_NORESERVE flag is ignored.
0050
0051
0052 How It Works
0053 ============
0054
0055 The overcommit is based on the following rules
0056
0057 For a file backed map
0058 | SHARED or READ-only - 0 cost (the file is the map not swap)
0059 | PRIVATE WRITABLE - size of mapping per instance
0060
0061 For an anonymous or ``/dev/zero`` map
0062 | SHARED - size of mapping
0063 | PRIVATE READ-only - 0 cost (but of little use)
0064 | PRIVATE WRITABLE - size of mapping per instance
0065
0066 Additional accounting
0067 | Pages made writable copies by mmap
0068 | shmfs memory drawn from the same pool
0069
0070 Status
0071 ======
0072
0073 * We account mmap memory mappings
0074 * We account mprotect changes in commit
0075 * We account mremap changes in size
0076 * We account brk
0077 * We account munmap
0078 * We report the commit status in /proc
0079 * Account and check on fork
0080 * Review stack handling/building on exec
0081 * SHMfs accounting
0082 * Implement actual limit enforcement
0083
0084 To Do
0085 =====
0086 * Account ptrace pages (this is hard)
